Power System Security Assessment: Application of Learning Algorithms

Power System Security Assessment
Application of Learning Algorithms
Christian Andersson
Licentiate Thesis
Department of Industrial Electrical Engineering
and Automation
Department of
Industrial Electrical Engineering and Automation
Lund University
Box 118
SE-221 00 LUND
SWEDEN
http://www.iea.lth.se
ISBN 91-88934-39-X
CODEN:LUTEDX/(TEIE-1047)/1-90/(2005)
c Christian Andersson, 2005

Printed in Sweden by Media-Tryck
Lund University
Lund, 2005
ii
School of Technology and Society
Malmö University
iv
Abstract
The last years blackouts have indicated that the operation and control of
power systems may need to be improved. Even if a lot of data was available,
the operators at different control centers did not take the proper actions in
time to prevent the blackouts. This depends partly on the reorganization of
the control centers after the deregulation and partly on the lack of reliable de-
cision support systems when the system is close to instability. Motivated by
these facts, this thesis is focused on applying statistical learning algorithms
for identifying critical states in power systems. Instead of using a model of
the power system to estimate the state, measured variables are used as input
data to the algorithm. The algorithm classifies secure from insecure states
of the power system using the measured variables directly. The algorithm is
trained beforehand with data from a model of the power system.
The thesis uses two techniques, principal component analysis (PCA) and
support vector machines (SVM), in order to classify whether the power sys-
tem can withstand an (n−1)-fault during a variety of operational conditions.
The result of the classification with PCA is not satisfactory and the technique
is not appropriate for the classification problem. The result with SVM is
much more satisfying and the support vectors can be used on-line in order to
determine if the system is moving into a dangerous state, thus the operators
can be supported at an early stage and proper actions can be taken.
In this thesis it is shown that the scaling of the variables is important for
successful results. The measured data includes angle difference, busbar volt-
age and line current. Due to different units, such as kV, kA and MW, the data
must be preprocessed to obtain classification results that are satisfactory. A
new technique for finding the most important variables to measure or super-
vise is also presented. Guided by the support vectors the variables which has
large influence on the classification are indicated.
v
vi
Contents
Abstract iv
Preface 1
Acknowledgement . . . . . . . . . . . . . . . . . . . . . . 1
1 Introduction 3
1.1 Previous work . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Objectives and contributions of the thesis . . . . . . . . . . 6
1.3 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Power system control 9

2.1 Power system stability . . . . . . . . . . . . . . . . . . . . 9
2.2 Power system operation . . . . . . . . . . . . . . . . . . . . 12
3 Data set 15
3.1 Data set structure . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 The NORDEL power system . . . . . . . . . . . . . . . . . 16
3.3 NORDIC32 . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4 The 2k factorial design . . . . . . . . . . . . . . . . . . . . 21
3.5 Dimension criterion . . . . . . . . . . . . . . . . . . . . . . 22
3.6 Training set . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.7 Classification set . . . . . . . . . . . . . . . . . . . . . . . 24
3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4 Principal component analysis 27

4.1 Principal component analysis . . . . . . . . . . . . . . . . . 27
vii
4.2 Number of principal components . . . . . . . . . . . . . . . 29
4.3 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3.1 Scaling to keep the physical properties . . . . . . . . 31
4.4 Classifying power system stability with PCA . . . . . . . . 32
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5 Support vector machines 37

5.1 Linear support vector machines . . . . . . . . . . . . . . . . 37
5.1.1 Maximal margin optimization . . . . . . . . . . . . 38
5.1.2 Soft margin optimization . . . . . . . . . . . . . . . 40
5.2 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2.1 Scaling to keep the physical properties . . . . . . . . 43
5.2.2 Scaling of unit stator current after capacity . . . . . 44
5.2.3 Variables . . . . . . . . . . . . . . . . . . . . . . . 44
5.3 SVM training . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.4 Classifying with a hyperplane . . . . . . . . . . . . . . . . 47
5.5 Reducing variable sets . . . . . . . . . . . . . . . . . . . . 51
5.5.1 Different variable setup . . . . . . . . . . . . . . . . 52
5.6 Implementation . . . . . . . . . . . . . . . . . . . . . . . . 72
5.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6 Conclusions 75
6.1 Summary of the results . . . . . . . . . . . . . . . . . . . . 75
6.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Bibliography 77
viii
Preface
During the final year of my B.Sc. studies in 2001 I was introduced to the
challenges related to planning and operation of power system by my current
PhD-advisor Bo Eliasson. I was studying computer and electrical engineer-
ing at Campus Norrköping, Linköping Institute of Technology. When plan-
ning my undergraduate project I contacted Daniel Karlsson at ABB Tech-
nology Products and came to work with wide area protection system against
voltage collapse. The project was supervised by Bo Eliasson at Malmö Uni-
versity. It was during my undergraduate project I became aware of the special
problems with multi input multi output system (MIMO), many of the power
systems today are probably the largest MIMO system ever built by mankind.
In the summer of 2002 I was appointed a PhD position at Malmö University
and was accepted to the post-graduate study program at Lund Institute of
Technology.
1
2
Acknowledgement
First of all, I would like to thank my supervisors, Bo Eliasson for introducing

me to the interesting problems in power system control and for his support
and advice. Gustaf Olsson for accepting me as a Ph.D. student at his depart-
ment and for his encouragement and optimism. Olof Samuelsson who leads
the electric power system automation research group at IEA, the guidance,
support and comments on my work have been very important.
I am also very grateful to Sture Lindahl for support and many interesting
discussions. Jan Erik Solem for introducing me to pattern recognition. The
constant guidance, support and unlimited help has helped me a lot, this thesis
had not been made without you. Lars Lindgren for help with the ARISTO
simulator and all the discussions about power systems we have had, I have
learned a lot from you.
I would like to thank the graduate students and researchers of the Indus-
trial Electrical Engineering and Automation at Lund Institute of Technology
and Applied Mathematics Group at Malmö University, especially to: Erik
Alpkvist, Yuanji Cheng, Pär Hammarstedt, Adam Karlsson, Niels Chris-
tian Overgard, Fredrik Nyberg and Markus Persson for providing a creative,
happy and sometimes confused atmosphere.
Furthermore, I would like to thank Gun Cervin and Yvonne Olsson at
Kranenbiblioteket, Malmö University for their excellent service. I am also
grateful to Jan Erik Solem, Olof Samuelsson and Christian Rosén for proof-
reading this thesis.
I would like to thank my family and friends, especially my two younger
sisters for their patience, even though I am always teasing them, (Josefin, jag
vet att jag är ditt elände). Finally, Cissi, you´re the best.
Malmö, August 21, 2005

Christian Andersson
Chapter 1
Introduction
In August and September 2003 three major blackouts occurred in N.E. North
America, S. Scandinavia and Italy. About 110 million people were affected
by these events. The interruption time ranged from a couple of hours to days.
The reason why these blackouts occurred were partly due to the lack of sup-
portive applications when the system is close to instability. If the power
system is close to the stability limit, actions must be taken by system opera-
tors to counteract a prospective blackout. Today a large problem for system
operators is to identify critical states since there are thousands of parameters
which describe and affect the state.
The deregulation of the power markets has contributed to the fact that the
power systems today are frequently operated close to stability limit. This is
due to a larger focus on economical profits instead of system security and
preventive maintenance.
The last three years the consumption of electrical power in the world has
increased with 3 %, each year. That corresponds to investments of 60 bil-
lion USD each year in production and transmission equipment to match the
increased consumption. If the trend continues actions must be taken to main-
tain system stability.
The latest blackouts and deregulated power markets contribute to an in-
creased requirement of Wide Area Protection Schemes (WAPS). The WAPS
will make it possible to change the operation criteria from: ”The power sys-
tem should withstand the most severe credible contingency” to ”The power
system should withstand the most severe credible contingency, followed by
protective actions from the WAPS.”
3
4 Chapter 1. Introduction
In this thesis classification of power system stability defines a method to

separate secure from insecure states in power systems. Due to the complex-
ity of power system with the large amount of states and the large amount
of variables that describe the states it is difficult to estimate the stability
of power systems. The purpose with this thesis is to examine methods for
classifying power system stability with techniques from pattern recognition.
The methods use the measured variables, such as busbar voltage, line current
and active current of generators, of the power system to decide the stability
of the power system. The classification algorithm which is developed with
techniques for pattern recognition is trained from a data base. The data base,
denoted training set, consists of measurements from different states of the
power system. With the algorithm it should be able to estimate the stability
of all the states of the power system may operate in. Therefore it is important
that the training set properly describes the variations of the power system. It
must be considered that if the training set does not show enough variability,
the algorithm is useless. A properly trained algorithm can separate the train-
ing set into two parts. One consists of stable states and the other of unstable
states. The algorithm can then be used to decide the class of unknown states.
This should be used as a supportive tool for the system operators to faster
form an opinion of the actual state of the power system. This could also
make it easier to decide what needs to be done by the system operators to
avoid a blackout. Compared with the WAPS the classification method should
prompt system operators at control centers to take preventive actions before
the WAPS will act.
The work with this thesis started with the development of the training set.
Thereafter Principal Component Analysis (PCA) was applied to the classifi-
cation problem. The result with PCA was not satisfactory. Therefore Sup-
port Vector Machines (SVM) was used to find a solution of the classification
problem. SVM calculates a hyperplane that separates a data set from their
predefined classes. The SVM method is more adaptable to the classification
problem in this thesis than PCA and the result is much more satisfying.
1.1. Previous work 5
1.1 Previous work
This thesis deals with the problem of classification of power system stabil-
ity. An overview of power system stability and design can be found in the
standard textbooks [13] and [22]. In [23] the CIGRE Study Committee 38
and the IEEE Power System Dynamic Performance Committee gives a back-
ground and overview of power system stability with the focus on definition
and classification. A nice introduction to power system voltage stability can
be found in [34], which has a detailed description of reactive power com-
pensation and control. These books and reports gives a good overview and
background to power system design and control.
Previous work on power systems stability has been focused on identifica-
tion and finding methods to prevent instability. In [6] and [26] the problem
with voltage instability is analyzed, these two theses are a part of the previ-
ous work that is the inspiration to this thesis. Another work that has been
a great inspiration to this thesis is the research by Louis Wehenkel which is
summarized in [39]. Wehenkel uses the same approach to the stability prob-
lem as in this thesis but he uses decision trees instead of principal component
analysis and support vector machines.
The knowledge about separation with a hyperplane has been known for
quite long time, but the calculation of the hyperplane has caused problems.
In 1992 Support Vector Machines (SVM) was introduced by Vapnik, see
[37]. It is described in detail in [36]. The SVM calculate a hyperplane with
the maximal margin between the two classes. This is done using a global
optimization problem on quadratic form, which can easily be solved.
The last years SVM has been used to find methods to solve many classi-
fication and regression problems. In [16] and [31], SVM has been applied
to face detection with successful results. Text categorization has also been
improved with SVM, see [18]. In [19], windows are detected from pictures
of city scenes. The author has found one publication which deals with clas-
sification of power system stability using SVM, see [30]. In the article clas-
sification of power system stability is made with nonlinear SVM and neural
networks. The data set which is used has an impressive size, and the result
of the classification is very successful.
1.2 Objectives and contributions of the thesis
The main objective of this thesis is to investigate if it is possible to use sta-

tistical learning algorithms for classification of power system stability. The
objectives of this thesis are listed below.
• Find a method which makes it possible to classify the stability of a

power system from measurements without calculating the state of the
power system.
• Adjust the classification method for the application to power system

stability.
• Investigate if the classification method can be implemented in real-

time so that operation of power systems can be simplified and also
improved.
People with different background should be able to read and comprehend the
thesis. This means that some readers will consider some parts of the thesis as
obvious. However, if the reader is not oriented in power system fundamentals
an introduction can be found in [13] and [22]. The main contributions of the
thesis are listed below.
• The power system stability is classified with a hyperplane, which is

calculated with Support Vector Machines (SVM). The SVM offers a
classification method with the largest distance between the two classes.
• Methods to improve the classification by using an appropriate scaling

algorithm. Investigations of which variables should be used to repre-
sent the power system in order to improve the classification result.
• Methods to reduce the number of measured variables are introduced.

Thus the classification can be made by a subset of the measurable vari-
ables without losing the number of correct classifications.
• Method for handling the difficulties of classification with statistical

learning algorithm with a mixture of continuous and binary variables.
1.3. Outline of the thesis 7
1.3 Outline of the thesis
The thesis consists of two parts, the design of the data set and two different
methods applied to the classification problem.
In the first part, Chapter 2, an overview of power system stability and op-
eration is given. Furthermore, the contribution of this thesis to the power sta-
bility problem is described. In Chapter 3, a short summary of the NORDEL
power system and the operation rules is given. Thereafter, an introduction to
design of experiments with factorial design is given. Finally, the design of
the data set is described.
The second part consists of Chapter 4 and Chapter 5. In Chapter 4 the
unsupervised learning method principal component analysis is applied to the
classification problem and in Chapter 5 the supervised learning method of
calculating a hyperplane with support vector machines is applied to the clas-
sification problem. Chapter 6 summarizes the results and suggests future
work.
1.4 Publications
The thesis is based on the following publications:

[1] C. Andersson, J E. Solem and B. Eliasson. Classification of power
system stability using support vector machines. In IEEE Power En-
gineering Society General Meeting. San Francisco, California, USA,
June 12-16. 2005.
[2] L. Lindgren, B. Eliasson and C. Andersson. Damping of power system

oscillations using phasor measurement units and braking resistors. In
IFAC proceedings of reglermöte. Gothenburg, Sweden, May 26-27.
2004.
[3] B. Eliasson and C. Andersson. New selective control strategy of power

system properties. In IEEE Power Engineering Society General Meet-
ing. Toronto, Ontario, Canada, July 13-17. 2003.
[4] B. Eliasson and C. Andersson. New selective control strategy of the

voltage collapse problem. In CIRED 17th International Conference
on Electricity Distribution. Barcelona, Spain, May 12-15. 2003.
Chapter 2
Power system control
This chapter consists of a brief overview of power system operation and the
security limits that are taken to maintain system stability in stressed situa-
tions.
2.1 Power system stability
The stability of power systems is defined as: "The ability of an electric power
system, for a given initial operating condition, to regain a state of operating
equilibrium after being subjected to a physical disturbance, with most sys-
tem variables bounded so that practically the entire system remains intact",
[23]. This means that some faults should not cause any instability problems.
Figure 2.1 shows different stability states. Due to the complexity of power
systems, the different instabilities affect each other. A blackout scenario
often starts with a single stability problem but after the initial phase other
instability problems may occur. Therefore the actions in order to prevent a
blackout in stressed situations may be different. For a detailed description of
the different stability scenarios see [23]. Due to mainly economical factors
there is a limit of which faults the system should withstand and remain in
stability.
The synchronous NORDEL system (Sweden, Norway, Finland and Sealand)
is dimensioned to fulfill (n−1) criterion, which means that the system should
withstand a single fault, e.g. short circuit of busbar or disconnection of nu-
clear unit, see [1]. In Figure 2.2 the different operative states are illustrated
and how the power systems operation state can change due to a disturbance.
9
10 Chapter 2. Power system control
Power System
Stability
Rotor Angle Frequency Voltage

Stability Stability Stability
Small-Disturbance Small-Disturbance Large-Disturbance Small-Disturbance

Angle Stability Angle Stability Voltage Stability Voltage Stability
Short Term Short Term Long Term Short Term Long Term
Figure 2.1: Classification of power system stability. Adapted from [23] .
In the secure state where the operating point of a power system should nor-
mally be it is possible to maintain system stability after a (n − 1) fault. It is
very important to have tools to verify the operating point of the system. In
this thesis a new method to improve the classification of the power system
stability is introduced. The method also makes it possible to deicide how
close the system is to the dividing line between the secure and insecure state.
However, faults which are included in the (n − 1) criterion sometimes
cause blackouts. In Sweden this happened Dec. 27, 1983 when a broken dis-
connector caused a short circuit of a busbar which lead to a blackout where
60 % of the load was disconnected from the system. The time to restore
normal operation was about 7 h. In the N.E. of North America on Aug. 14,
2003, 50 million people were affected by a blackout caused by a cascade
disconnection of transmission lines. The time between the disconnection of
the first transmission line and the blackout was 2,5 hours and the restoration
took almost 48 hours. In Italy in Sep. 28, 2003, 53 million people were
affected by a blackout which also was caused by a cascade disconnection of
transmission lines. The fault scenario was a bit more difficult to overview
than in USA and Canada. First a transmission line was disconnected due to
a short-circuit caused by a line sagging into a tree and then after 25 minutes
twelve other lines were disconnected within 2,5 minutes which caused the
blackout. The restoration time was 18 hours and the number of switching
operations that had to be made was about 3600.
2.1. Power system stability 11
PREVENTIVE STATE
NORMAL
SECURE
Maximize economy and minimize the effect of uncertain contingencies
Preventive control
RESTORATIVE ALERT
INSECURE
Resynchronization Tradeoff of
Load pickup preventive vs
corrective control
Emergency control
(corrective)
IN EXTREMIS EMERGENCY
Protections
ASECURE
Partial or total Overloads

service Undervoltages
interruption Split Underfrequency...
Load shedding Instabilities
Control and / or protective actions
Foreseen or unforeseen disturbance
Figure 2.2: The operating states and transitions for power systems. Adapted
from [12] .
In Figure 2.2 the state during abnormal situations are described with the
insecure and asecure states. If a power systems operating point moves from
the secure to the insecure state, actions must be taken by system operators
to return to the secure state. By the three blackouts in Sweden, N.E. North
America and Italy, it is shown that the preventive time interval can be be-
tween 50 seconds to 2,5 hours. From a system operation point of view it is
possible to take actions before entering the insecure state. This thesis has
focus on supportive information, before entering the insecure area. In the
asecure state, only automatic actions and control are realistic.
Blackouts have occurred and will happen in the future, but the objective
from the power industry is to increase the time between the blackouts and
also minimize the number of people which are affected. By e.g. support-
ive applications for the operators based on new technology, the prevention
against blackouts can be improved. By increased automation the imbalances
between production and consumption can be quickly alleviated. Also the
black-start capability after a blackout can be improved.
2.2 Power system operation
In power system control today the operating point is validated by a state

estimator and tools to calculate system collapse limit. There are different
software tools developed for this. In [35] the experience with implementing
a voltage security assessment tool at B.C. Hydro, Burnaby, Canada is de-
scribed. In Sweden today SPICA is used to calculate the voltage collapse
limits. SPICA is developed by Swedish National Grid (Svenska Kraftnät)
and uses an estimate to calculate the voltage collapse limits by load-flow
calculations. Mainly by increasing the load in sensitive parts of the system
and exposing them for a selection of (n − 1) faults. The predictor proposes
changes in such as,
• Transfer limits
• Increased production
• Decreased export by HVDC,
if the estimated operating point does not fulfill the (n − 1) condition.
The limitations with the state estimators are that they use a model of the
power system. There is always a difference between the model and the real
2.2. Power system operation 13
system. It is also common to use load-flow calculations. In a load-flow

calculation the dynamic part of the elements in the model are neglected e.g.
load dynamic and also often voltage regulators and current limiters. The
load-flow solver calculates an operating point where there is balance in active
and reactive power. There is a difference between load-flow calculation and
dynamic simulation. When the operating condition changes, e.g. by a short-
circuit of a busbar, the load-flow solver calculates the operating point after
clearance of the fault. The dynamic simulation also analyzes the effect by
e.g. automatic equipment. Hence the operating point after the transients
are settled will be different compared with a load-flow calculation. This
can result in that the operating point calculated by load-flow is stable but if
the scenario is analyzed by dynamic simulation a blackout can occur in the
transient phase.
In this thesis, dynamic simulations are used to develop classification meth-
ods which are used to estimate the bound between the secure and the insecure
state, see Figure 2.2. Instead of using a traditional state estimator, statistical
learning algorithms are used to calculate an algorithm which uses measured
variables, e.g. busbar voltage, angle difference and line current to classify
the stability.
Chapter 3
Data set
In this chapter the data set for verifying the statistical learning algorithms
applied on classification of power system stability is described. With the
algorithm is should be possible to estimate the class for all states in the power
system. It is therefore important that data set properly describe the variations
of the system. It must be considered that if the data set does not show enough
variability, the SVM cannot be successfully applied.
To create a function for classification with a learning machine with the best
conditions, the space of operation should be covered by the data set. In e.g.
chemical processes, design of experiments are used to cover the operation
space with a limited number of experiments, for details see [10].
The data set used to verify the classification algorithms has not been de-
signed with experimental design since it would become very large. The time
it takes to prepare the simulator and collect the data are not covered in this
project. Instead, the knowledge about the daily operation of the NORDEL
system and the problems that has caused stressed operation situations has
been taken into account to create the data set.
3.1 Data set structure
In this thesis the data set is organized as a matrix, see Table 3.1. A row in the
matrix, denoted with an example, is the state of the power system contain-
ing the measured variables. The label vector consists of the binary code
for each example. If the example sustains the dimension criterion (n − 1),
see Section 3.5, has the binary code +1, otherwise -1.
15
16 Chapter 3. Data set
Table 3.1: Description of how the data set is organized, the label vector con-
sists of the binary code to the corresponding examples.
Variables Label
Example 1 +1
.. ..
. .
Example l -1
3.2 The NORDEL power system
In this presentation of NORDEL, Iceland and Jutland (Denmark) are ex-

cluded, for details see Figure 3.1. The synchronous system consisting of
Finland, Norway, Sealand (Denmark) and Sweden is presented. The installed
capacity was 93 GW on Dec. 31, 2003, the maximum load was 58 GW, mea-
sured on Jan. 15, 2003. In Table 3.2 the distribution of the installed capacity
per country is listed. The information about the NORDEL power system
published here is obtained from [4].
In Sweden, a large amount of hydro power is located in the northern part
and nuclear power is located in the southern part. Because of a larger load de-
mand in the south of Sweden the power flow normally is from north to south.
The latest blackouts in Sweden (1983 and 2003) occurred because of reduced
capacity in the transmission system, see [3] and [7]. The NORDEL system
data is not used here, but instead a simplified model called NORDIC32 is
used. This simplified model should be compared with the NORDEL power
system in order to give guidelines for the design of the data set. To classify
the stability of a power system, traditionally bus voltage and system fre-
quency have been used. In Sweden the transfer of active power on roughly
10 selected transmission lines are used as guidance for the system stabil-
ity. The transmission system is divided into 4 cut-sets, called Snitt 1 − 4.
Snitt 4 measures the sum of active power of the 5 transmission lines that
connect the south of Sweden with the rest of the system, see Figure 3.1. Op-
eration aims at keeping the transport of active power through Snitt 4 below
a certain threshold to retain system stability from the most severe credible
contingency.
3.2. The NORDEL power system 17
Figure 3.1: The NORDEL main grid, 2000, The cut-sets, Snitt 1 to 4, are
marked.
Table 3.2: Installed capacity in NORDEL on 31 Dec. 2003, MW.

Denmark Finland Norway Sweden
Hydro power 11 2 978 27 676 16 143
Nuclear power - 2 640 - 9 441
Other thermal power 9 704 11 225 305 7 378
Renewable power 3 115 50 100 399
3.3 NORDIC32
The NORDIC32 (N32) system is used for the study [5]. It is an equivalent or
simplified model of a transmission system with similarities to the Swedish
power system, see Figure 3.2. The generation consists of approximately
45 % nuclear, 45 % hydro and 10 % other thermal power. The installed
capacity is 17 GW, see Table 3.3. The system has 32 substations, 35 units,
50 branches and the number of measurable variables are 653. To adjust the
spinning reserve and speed droop the relation between N32 and NORDEL
are settled to approximately 1:4.
In order to produce the data set, the platform described below was used.
Since 1996 the simulator ARISTO (Advanced Real-time Interactive Sim-
ulator for Training and Operation) has been used for operator training at
Swedish National Grid (Svenska Kraftnät). The simulator has unique fea-
tures for stability analysis in real time [9]. Not only that, the simulation
proves to be extremely robust during abnormal operating conditions. Using
ARISTO the operators obtain a better understanding of the phenomena that
may cause limitations and problems to power transmission and distribution
systems.
The load models in ARISTO cover the load dynamics (short and long
term) and load characteristics (voltage and frequency dependence) very well.
Transmission lines are represented with pi-equivalents. The generation units
are represented by synchronous generators and turbine governors, either ther-
mal or hydro. The protection equipment implemented in ARISTO are
• Line distance protection
• Line over-current protection

3.3. NORDIC32 19
Table 3.3: Installed capacity in NORDIC32, MW.

Finland Norway Sweden
Hydro power 4 605 1 520 4 740
Nuclear power - - 4950
Other thermal power - - 1 080
• Unit over/under frequency protection
• Unit over/under voltage protection
and the automatic equipment
• Load disconnection due to under-frequency
• Shunt capacitor/reactor connect/disconnect due to under/over voltage.
The detailed representation of power system models in ARISTO results in

a more realistic data set than if a simple load-flow solver had been used.
The difference between load-flow calculations and dynamic simulations is
that load-flow solvers calculate an operation point for the given conditions,
whereas dynamic simulations calculate a time-simulation where the oper-
ating point is calculated for each time step. The advantage with dynamic
simulations is that it is possible to analyze the influence of automatic protec-
tion, control equipment and dynamic models on the power system. In this
study a three-phase short-circuit on a busbar is analyzed. With the dynamic
simulation it is possible to take into account what has happened during the
transients. With a load-flow solver it is only possible to analyze if there
is balance in active and reactive power after the disconnection of the fault.
Therefore, the data set is more realistic with the dynamic simulation.
The main additives to the simulator consist of a multi-user communication
link between the Application Programmable Interface (API) of ARISTO and
Matlab/Simulink. This function is used to collect the data set.
Figure 3.2: The NORDIC32 system.

3.4. The 2k factorial design 21
3.4 The 2k factorial design
Factorial designs are often used in designs of experiments. The purpose

is to cover most of the variation in the process with a limited number of
experiments, see [10] and [29]. The elements in the design consist of factors
and response. Response is the output of the process. Factors are the elements
in the process that influence the output. The factors can be both controllable
and uncontrollable.
For classification of power system stability the label of the example is de-
cided according to the Nordel grid planning rules for the Nordic transmission
system, for details see [1]. Therefore the factorial design is applied with the
following factors
• Production
• Transmission capability
• Consumption
and responses
• Bus voltages
• Angle differences
• Line current
• System frequency
• Load disconnection from system protection.

The purpose with factorial design is to define a high and a low level for
every factor. In 23 design the factors form a cube where the examples shall
be chosen from the corners of the cube, see Figure 3.3. If the factorial design
is applied to the N32 system the number of simulations will be very large.
Taken into account that the number of units are 35, branches 50 and loads
59 the simulation time will be very long. Approximations can of course be
made. Using only nuclear power, a data set could be designed with factorial
design. The number of substations with nuclear power are 5, thus the number
of examples in the data set would be 32. The problem is to design a data set
that also has a variation in production of hydro units, power flow, and load,
Figure 3.3: Illustration of a factorial 23 design.
with a limited number of examples. Therefore factorial design is not used

to design the data set. Instead, knowledge from operation of the NORDEL
system with already known limits has been used in the design of the data set.
3.5 Dimension criterion
According to Nordic grid instructions, the Nordel system must fulfill the
(n − 1) criterion, see [1] and [2]. An example is considered acceptable if the
power system can withstand any single failure. The system has withstood a
contingency if:
• There is no load shedding, due to under frequency.
• The final post-fault voltage profile is acceptable.
• The final post-fault frequency is acceptable.
The contingencies are e.g. disconnection or short-circuit of a line, load,

shunt or busbar. After fault clearance the voltage profile is acceptable if
all except one busbar are within 10 % of the normal operating voltage. The
system frequency is accepted if it does not drop below 49.0 and exceeds 49.5
Hz after 30 seconds. The upper level demand on the frequency is 52.0 Hz
and it must be lower than 50.6 Hz after 30 seconds.
3.6. Training set 23
To verify the (n−1) criterion for the examples, a three-phase short-circuit

at all busbars, one at a time, is used to decide the binary code according to
the rules above. This is done for each of the training examples. The fault
is cleared 80 ms after the fault appeared, see [13]. In practice this is done
in ARISTO by first applying the fault to the example, then after 80 ms the
fault is cleared, finally after 240 s it is considered that the transients are
approximately settled and the simulation is stopped. The binary code is then
decided for the example.
Because N32 is an equivalent, with reduced topology, the (n − 1) criterion
is not fulfilled by most of the examples in the data set. The data set is split
into three classes:
1. Examples where a single fault leads to system black-out.
2. Examples where a single fault leads to black-out in two or three sub-

stations.
3. Examples which fulfill the (n − 1) criterion.
Therefore a method shall be developed to identify class 1 and separate

class 1 from class 2 and 3.
3.6 Training set
As training data, four base generation patterns are used.
• NN, all nuclear power at maximum production.
• NO, all nuclear power at maximum production on the west coast, no

nuclear power on the east coast.
• ON, all nuclear power at maximum production on the east coast, no

nuclear power on the west coast.
• OO, no nuclear power on-line.
The first step is to find the minimum load level for the four generation pat-
terns, by scaling the active power part of the load on a common system base.
To create balance in the system at the minimum level the production of hy-
dro power is corrected. As little hydro power as possible should be used to
create balance in the minimum case. In order to achieve training data with
the characteristics of the NORDEL power system the following dimension
criteria are used for all examples.
• The system frequency is within 50 ± 0,01 Hz.
• The voltage profile is adjusted according to [24], where the normal op-
eration voltage on the 400 kV level is 400-415 kV, with the minimum
accepted voltage 395 kV and the maximum accepted voltage 420 kV.
• The spinning reserve in the system is 750 MW, divided on three hydro
power units. It corresponds to the largest unit in N32.
• The ratio between N32 and NORDEL is approximately 1:4. Taken this
into account the speed droop is tuned to about 1500 MW/Hz for all
models. The Nordel system (Sweden, Finland, Norway and Sealand)
it is 6000 MW/Hz, see [2].
When the minimum load level is achieved, a new model is created by increas-
ing the load level by 5 % increments from the base load profile. Spinning
reserve and the speed droop is tuned so that all models have similar charac-
teristic. For the 4 generation patterns the load level is increased with 5 %
until the maximum load level is identified. The total number of models in
the training data is 30, with a consumption between 3 280 and 13 570 MW.
This is done in order to extend the generation/transmission patterns as much
as possible with a limited number of training models. The generation and
consumption of reactive and active power for the 30 examples in the training
set are listed in Table 3.4. The label for the examples are also listed here.
Table 3.4 shows that the increased consumption for the three different base
patterns, NN, NO, ON, does not imply a more stressed state.
3.7 Classification set
The classification algorithm is developed with a training set and validated

with a classification set. To show the performance of the algorithm. The
classification set consists of 16 examples. These are created from 16 of the
examples in the training data that have been manipulated e.g. increased load
factor and disconnection of single line or unit. Thus, an example which
3.7. Classification set 25
Table 3.4: The system generation and consumption for the examples in the
training set, MW and MVar.
Name PGeneration Pload QGeneration Qload Label
NN1 7 489 7 329 1 550 3 359 -1
NN2 8 869 8 752 975 3 359 1
NN3 9 968 9 846 1 244 3 359 1
NN4 10 545 10 392 907 3 359 -1
NN5 11 136 10 940 1 019 3 359 -1
NN6 11 707 11 487 970 3 359 1
NN7 12 463 12 034 2 386 3 359 -1
NN8 12 948 12 581 1 640 3 359 1
NN9 13 545 13 128 2 019 3 359 1
NN10 14 075 13 566 2 832 3 359 -1
NO1 4 541 4 376 1 126 3 359 1
NO2 5 667 5 470 857 3 359 1
NO3 6 826 6 564 998 3 359 1
NO4 7 418 7 111 1 199 3 359 -1
NO5 8 021 7 658 1 440 3 359 -1
NO6 8 630 8 205 1 581 3 359 -1
NO7 9 266 8 752 2 408 3 359 -1
ON1 4 468 4 375 915 3 359 1
ON2 5 556 5 470 873 3 359 1
ON3 6 674 6 564 798 3 359 1
ON4 7 243 7 111 780 3 359 -1
ON5 7 819 7 657 777 3 359 -1
ON6 8 405 8 205 908 3 359 1
ON7 8 998 8 752 1 041 3 359 1
ON8 9 608 9 299 1 382 3 359 1
ON9 10 203 9 846 1 580 3 359 1
ON10 10 835 10 393 2 262 3 359 1
OO1 3 428 3 282 760 3 359 -1
OO2 4 639 4 376 1 075 3 359 -1
OO3 5 905 5 470 2 254 3 359 -1
fulfills the criterion in the training set is manipulated so that it does not fulfill
the criterion in the classification set. The binary code for the examples is
decided with the same criterion as for the training set.
3.8 Conclusions
In this chapter the data set that will be used to verify the classification meth-
ods in Chapter 4 and 5 described. The methods to design the different ex-
amples and the dimension criterion for deciding the binary code are pre-
sented. The detailed protection equipment in ARISTO, adjustment of spin-
ning reserve, speed-droop, transient analysis and comparison to Nordel grid
instructions improves the quality of the data set. The ARISTO simulator is
used instead of a normal load flow solver, which improves the quality.
Chapter 4
Principal component analysis
In this chapter an introduction to Principal Component Analysis (PCA) will

be given. PCA is one of the most popular unsupervised learning methods.
The main reason is that the method reduces the dimension of the data set.
With PCA the variables in the data set are projected to a lower dimensional
space described by the principal components, see [14], [15] and [20]. Thus
the characteristics of a multivariate process can hopefully be shown in lower
dimensions. The projection of the data set on the principal components can
give an unexpected relationship between the process and the variables.
Today in supervision and control of power systems the main problems
are not to measure and collect data. The Supervisory Control And Data
Acquisition (SCADA) system gives this possibility. The main problem is to
identify examples which do not fulfill the (n − 1) criterion. Furthermore, it
might not be necessary to use all measured variables from the power system
for the identification. Which variables that can be reduced is an important
question to investigate.
4.1 Principal component analysis
PCA was introduced by Pearson in 1901 [32]. With PCA the dimension of
the data set is reduced from Rd to Rn , where n ≤ d. The principal com-
ponents are orthogonal in Rn . The principal components are the directions
in the data set with the highest variance, see Figure 4.1. An introduction to
how the principal components are calculated will be given below.
27
28 Chapter 4. Principal component analysis
* *
W
1
* *
* * *
* *
* *
*
* * * *
* *
* *
* * *
Figure 4.1: The first linear principal component of a data set. The line mini-
mizes the total squared distance from each point to its orthogonal
projection onto the line.
Denote the data set

Xl×d ,
where X is the data set stored as l examples with d variables. Let x =

(x1 , . . . , xd ) be the mean value for all variables and,
X̂ = [x1 − x1 , . . . , xd − xd ] ,
where xi , i = 1, . . . , d are the columns in X.
The covariance matrix of X̂, see [25], is
1
Σ= X̂> X̂ .
l−1
The principal components of X̂ can be calculated by solving the eigenvalue

problem,
ΣΦ = ΛΦ ,
4.2. Number of principal components 29
where Φ is an orthogonal matrix and Λ is a diagonal matrix containing the

eigenvalues. The eigenvectors to X̂> X̂, Φ can also be calculated with Sin-
gular Value Decomposition (SVD) of X̂, as
X̂ = USV> ,
where Ul×l and Vd×d are orthogonal matrices and Sl×d is a diagonal matrix
containing the singular values. The first principal component, the direction in
the data set with largest variance captured with a linear function, is the eigen-
vector with the largest corresponding singular value. The SVD-factorization
is usually defined with the singular values in descending order.
In [15] and [11] the calculation of principal components are denoted with
X̂ = TP + E , (4.1)
where Tl×n are called score vectors, Pd×n are called loading vectors, n is
the number of principal components and E is the residual or variance in the
data set not described by the principal components. T and P are calculated
as,
I = {i1 , i2 , . . . , id }
J = {i1 , i2 , . . . , in }
P = V> {I, J}
T = X̂P> ,
where V> {I, J} are all the rows and the n first columns of V > .
4.2 Number of principal components
How to choose the correct number of principal components is something that

is difficult to answer. There are some values that can be calculated to guide
the decision. If too many principal components are used the dimension of
the data set is not reduced enough. On the other hand, if too few principal
components are used, the residuals will be too large. For PCA the number
of components is usually chosen by guidance from the eigenvalues. In Fig-
ure 4.2 the singular values of X b are shown, it is sometimes denoted with
scree plot [20].
Singular value
0 i
0 1 2 3 4 5 6 7
Principal component
Figure 4.2: An example of a scree plot, showing the singular values.
There is a ”knee” in the plot when i = 3, therefore according to a popular

rule, the number of components should be 3, see [20]. In [15] two factors are
introduced for improving the choice of the number of principal components,
Hvalue and Herror . They are based on the residuals variance
d
! n !
X X 1
Hvalue = λi
λ
i=n+1 i=1 i
d
! n
!
X X 1
Herror = λi 1+ ,
λi
i=n+1 i=1
b > X.
where λi are the eigenvalues to X b The number of principal components
should be chosen so Hvalue and Herror are minimized. Both functions can
be used but it is often more appropriate to use H error because it is more
distinct, compared to Hvalue .
4.3 Scaling
Before calculating the principal components it is useful to preprocess the

data set. The different variables can have different range before preprocess-
4.3. Scaling 31
ing, e.g. the angle difference is between 0-90 degrees and the production of
active power difference between 0-600 MW. In practice, the examples in the
data set xi are stored as rows in a matrix. The columns z = (z 1 , . . . , zl )
of this matrix are scaled to a predefined interval. There are different meth-
ods to scale the variables in the examples e.g. mean-centering and univariate
scaling, see [11].
z−m
zunivariate = , where (4.2)
σ
l
1X
m = zi
l
i=1
v
u l
uX (zi − m)2
σ = t .
l−1
i=1
4.3.1 Scaling to keep the physical properties
To keep the physical properties of the system a new scaling method is in-
troduced, by scaling the variables group-wise. First, the voltage V and I are
calculated in p.u. [13] as
V I
Vp.u. = and Ip.u. = ,
Vbase Ibase
with a common system base Sbase . Which results in,
Sbase
Ibase = √ .
3Vbase
Then the variables are divided into groups, e.g. busbar V, load P, line I and
angle differences. The groups are scaled with the univariate function (4.2).
Instead of dividing with the standard deviation in every column, the columns
are divided with the largest absolute value found in groups of variables.
z−m
zgroup = , where (4.3)
Max
Max = The largest absolute value in a group of variables .
7
x 10
2
8000
1.8
7000
1.6
6000
1.4
5000 1.2
4000 1
0.8
3000
0.6
2000
0.4
1000
0.2
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
(a) (b)
Figure 4.3: (a) The scree plot for the first 30 singular values to the training set.
(b) The 30 first numbers of Herror for the training set.
4.4 Classifying power system stability with PCA
The principal components are calculated for the training set, see Section 3.6.
According to Figure 4.3 (a), the number of principal components should be
between six or seven if the ”knee” condition is used. In Figure 4.3 (b) the
number of principal components should be 29, if H error should be mini-
mized. The purpose with principal components, to reduce the dimension,
is not fulfilled according to Herror . In Figure 4.4 to 4.7 the score plots for
the five first principal components are presented. The two different scaling
methods from Section 4.3 are used. It is clearly shown in Figure 4.4 to 4.7
that it is impossible to linearly separate the examples in the training set with
different binary code using these five principal components. Experiments
have shown that 19 principal components are needed to linearly separate the
examples.
4.5 Conclusions
In this chapter the possibility to use principal component analysis to classify

the stability of the data set presented in Chapter 3 is analyzed. Due to the
complexity of power systems, with the large amount of variables that de-
scribe the state of the system, it is not possible to use PCA to separate the
4.5. Conclusions 33
40 40
30 30
Second Principal Component
Third Principal Component

20 20
10 10
0 0
−10 −10
−20 −20
−30 −30
−40 −40
−40 −30 −20 −10 0 10 20 30 40 −40 −30 −20 −10 0 10 20 30 40
First Principal Component First Principal Component
40 40
30 30
Fourth Principal Component
20
Fifth Principal Component 20
10 10
0 0
−10 −10
−20 −20
−30 −30
−40 −40
−40 −30 −20 −10 0 10 20 30 40 −40 −30 −20 −10 0 10 20 30 40
Figure 4.4: Score plot for the 5 first principal components. The examples are
marked with ”o” or ”*”, where ”o” has the binary code -1 and ”*”
has the binary code 1. The data set are scaled univariate, see (4.2) .
examples in the training set with different binary code. A method for separa-
tion in higher dimensions must be chosen to succeed with the classifications.
40 40
30 30

20 20
10 10
0 0
−10 −10
−20 −20
−30 −30
−40 −40
−40 −30 −20 −10 0 10 20 30 40 −40 −30 −20 −10 0 10 20 30 40
Second Principal Component Second Principal Component
40 40
30 30

Fifth Principal Component
20 20
10 10
0 0
−10 −10
−20 −20
−30 −30
−40 −40
−40 −30 −20 −10 0 10 20 30 40 −40 −30 −20 −10 0 10 20 30 40
Second Principal Component Third Principal Component
40 40
30 30
20 20
10 10
0 0
−10 −10
−20 −20
−30 −30
−40 −40
−40 −30 −20 −10 0 10 20 30 40 −40 −30 −20 −10 0 10 20 30 40
Third Principal Component Fourth Principal Component
has the binary code 1. The data set are scaled univariate, see (4.2) .
4.5. Conclusions 35
10 10
8 8
6 6
Second Principal Component

4 4
2 2
0 0
−2 −2
−4 −4
−6 −6
−8 −8
−10 −10
−10 −5 0 5 10 −10 −5 0 5 10
10 10
8 8
6 6
4 4
2 2
0 0
−2 −2
−4 −4
−6 −6
−8 −8
−10 −10
−10 −5 0 5 10 −10 −5 0 5 10
10 10
8 8
6 6
4 4
2 2
0 0
−2 −2
−4 −4
−6 −6
−8 −8
−10 −10
−10 −5 0 5 10 −10 −5 0 5 10
Second Principal Component Second Principal Component
has the binary code 1. The data set are scaled group-wise, see
(4.3) .
10 10
8 8
6 6

4 4
2 2
0 0
−2 −2
−4 −4
−6 −6
−8 −8
−10 −10
−10 −5 0 5 10 −10 −5 0 5 10
Second Principal Component Third Principal Component
10 10
8 8
6 6
4 4
2 2
0 0
−2 −2
−4 −4
−6 −6
−8 −8
−10 −10
−10 −5 0 5 10 −10 −5 0 5 10
Third Principal Component Fourth Principal Component
has the binary code 1. The data set are scaled group-wise, see
(4.3) .
Chapter 5
Support vector machines
This chapter is an introduction to Support Vector Machines (SVM), a new

generation learning system based on recent advances in statistical learning
theory. Due to the large amount of variables that describe a state and the
difficulties to estimate the stability of a state SVM is used as a classifica-
tion tool. The aim is to use support vectors for calculating a hyperplane
in high dimensions. This hyperplane separates the data with the maximum
distance between the classes. The separation can be both linear and non-
linear. The theory is applied to classification of power system stability. SVM
have been applied to real-world applications such as text categorizing, hand-
written character recognition, image classification and biosequence analysis.
5.1 Linear support vector machines
SVM were introduced by Vapnik in 1992 [36] and has received considerable
attention recently. The objective is to create a learning machine that responds
with a binary output to the test data. If the system sustains the dimension
criterion it is marked by +1 otherwise it is marked -1. This designation is here
after called the binary code. In order to train the machine, the training data
must properly describe the variations of the system. It must be considered
that if the training data do not show enough variability, the SVM cannot be
successfully applied.
37
38 Chapter 5. Support vector machines
5.1.1 Maximal margin optimization

Denote the training data
{xi , yi }, i = 1, . . . , l ,
where xi ∈ Rd are the input measurements and yi ∈ {−1, 1} are the corre-
sponding binary responses. Suppose that there exists a hyperplane,
x·w+b = 0 ,
with parameters (w, b), separating positive from negative examples. Let d+
and d- denote the shortest distance from the hyperplane to the closest posi-
tive/negative point respectively. The goal of the training is to find the hyper-
plane which best separates the training data. For a linear SVM this results in
a convex optimization problem, which has a global solution. The hyperplane
is therefore the plane in Rd , with the largest d+ and d-. This is a useful prop-
erty that other classification methods e.g. neural networks [33] do not have.
The constraints that the hyperplane must separate the points,

xi · w + b ≥ 0, for yi = +1
xi · w + b ≤ 0, for yi = −1
may be combined into
yi (xi · w + b) ≥ 0 .
The points which lie on the hyperplane H 1 : xi · w + b = −1 have the
perpendicular distance from the origin |b − 1|/kwk, and the points on the
hyperplane H2 : xi · w + b = 1 have the perpendicular distance from the
origin |b + 1|/kwk. Hence the margin is 1/||w||, see Figure 5.1.
The optimization problem

Minimize ||w||2
(5.1)
under the constraint di = yi (x · w + b) ≥ 1 ,
can be solved, by the following transformation,
X l X l
1
L(w, b, a) = ||w||2 − ai yi (xi · w + b) + ai , (5.2)
2
i=1 i=1
5.1. Linear support vector machines 39
H2
o
H1
w
o o
*
* o
* *
d
+
d
−
Figure 5.1: Linear separating hyperplane.
where a = (a1 , . . . , al ) are the Lagrange multipliers. The function is min-

imized with respect to w and b. The constraints from the gradient of the
optimization problem with respect to w and b are
 l
 X
 ∂L(w, b, a) = w −

 ai yi xi = 0
 ∂w
i=1
 l
X

 ∂L(w, b, a)

 = ai yi = 0 .
∂b
i=1
Substituting these in (5.2) gives

l
X l
1 X
L(w, b, a) = ai − ai aj yi yj xi · x j . (5.3)
2
i=1 i,j=1
P
The SVM maximizing (5.3) with respect to a i for the constraints ai yi = 0
and ai ≥ 0 gives the desired hyperplane. The optimization problem has a
global solution, which gives the maximum margin between the two classes.
The Karush-Kuhn-Tucker (KKT) [8] condition is used to determine b, the
distance from the origin to the hyperplane, see Figure 5.1.
According to the KKT condition

ai (yi (xi · w + b) − 1) = 0 .
This can be used to determine b for any i for which a i > 0. Every training
point xi has a corresponding ai . The points which correspond to ai > 0 are
called support vectors. These points lie on the hyperplanes, H 1 and H2
and affect the decision of the machine.
The classification data can be classified with the hyperplane using the sign
of the function
l
X
f (x) = x · w + b = ai yi xi · x + b . (5.4)
i=1
5.1.2 Soft margin optimization

The maximal margin optimization calculates a hyperplane with a consistent
hypothesis. The resulting distance between the two classes is always the
largest possible distance. In some classification problems it is not desirable
to use the maximal margin optimization. If an example in the training set
consists of misleading data the hyperplane will be affected by this. There-
fore soft margin optimization can be used to calculate a hyperplane, which
permit some degree of incorrect separation. This hyperplane represents a
simpler model than the possibly over-fitted hyperplane of the maximal mar-
gin classifier. The so called slack variables ξ are used to represent the in-
correct separation, see Figure 5.2. The hyperplane is calculated from the
following optimization problem, for details see [8].
2-Norm Soft Margin

Using the 2-norm for the slack variable, the optimization problem is

Xl

 2
Minimize ||w|| + C ξi2
(5.5)

 i=1
under the constraint di = yi (x · w + b) ≥ 1 − ξi ,
where ξ = (ξ1 , . . . , ξl ) are the slack variables. The variable C must be cho-
sen from a wide range of values. By using cross validation, leave-one-out
5.1. Linear support vector machines 41
H2
*j
ξj o
H1
ξi w
o o
oi *
* o
* *
d
+
d
−
Figure 5.2: Linear separating hyperplane, with slack variables.
test, or a separate validation test set the value of the parameter is decided to
obtain optimal performance. In this thesis C is set to 1000. The optimization
test has not been performed to decide the value of C, due to the limited num-
ber of training examples. For implementation in power system operation the
performance will be improved if C has the optimal value. The corresponding
Lagrangian for the 2-norm soft margin optimization problem is
l l l
1 CX 2 X X
L(w, b, ξ, a) = ||w||2 + ξi − ai yi (xi · w + b) + ai (1 − ξi ) ,
2 2
i=1 i=1 i=1
(5.6)
where a = (a1 , . . . , al ) are the Lagrange multipliers. The constraints from
the gradient of the optimization problem with the respect to w, ξ and b are
 l

 ∂L(w, b, ξ, a) X

 = w − ai yi xi = 0

 ∂w

 i=1
 ∂L(w, b, ξ, a)
= Cξ − a = 0

 ∂ξ

 l

 ∂L(w, b, ξ, a) X

 = ai yi = 0 .
 ∂b
i=1

l
X l l
1 X 1 X 2
L(w, b, a) = ai − ai aj yi yj xi · x j − ai . (5.7)
2 2C
i=1 i,j=1 i=1
The SVM maximizing (5.7) with respect to a i , for the constraints

P
ai yi = 0 and ai ≥ 0, gives the desired hyperplane.
According to the KKT condition b is determined with
ai (yi (x · w + b) − 1 + ξi ) = 0 ,
for any ai > 0.
1-Norm Soft Margin

Using the 1-norm for the slack variable, the optimization problem is
 l

 X

 Minimize 2
||w|| + C ξi
i=1 (5.8)

 under the constraint di = yi (x · w + b) ≥ 1 − ξi


ξi ≥ 0 .
The 1-norm soft margin is refered to as ”the Box Constraint”, where ξ i < C.
The difference between 1-norm and 2-norm is that ξ i is limited by C for 1-
norm, and not by 2-norm. The corresponding Lagrangian for the 1-norm soft
margin optimization problem is
X l X l X l
1
L(w, b, ξ, a, r) = ||w||2 +C ξi − ai yi (xi ·w+b−1+ξi )+ ri ξi ,
2
i=1 i=1 i=1
(5.9)
where a and r are the Lagrange multipliers. The function is minimized with
respect to w, ξ. The constrains from the gradient of the optimization problem
with the respect to w, ξ and b are
 l

 ∂L(w, b, ξ, a, r) X

 = w − ai yi xi = 0

 ∂w

 i=1
 ∂L(w, b, ξ, a, r)
= C − a i − ri = 0

 ∂ξi

 l

 ∂L(w, b, ξ, a, r) X

 = ai yi = 0 .
 ∂b
i=1
5.2. Scaling 43
l
X l
1 X
L(w, b, a) = ai − ai aj yi yj xi · x j , (5.10)
2
i=1 i,j=1
which is identical with the maximal margin. The only difference is the de-
termination if b with the KKT condition

ai (yi (x · w + b) − 1 + ξi ) = 0
ξi (ai − C) = 0 ,
for any ai > 0.
5.2 Scaling
As in Chapter 4 the data set is preprocessed before calculation of the hyper-

plane. In practice, the examples in the data set x i are stored as rows in a
matrix. The columns z = (z1 , . . . , zl ) of this matrix are scaled to a prede-
fined interval with a linear scaling [17],
2(z − Min)
zlinear = −1 (5.11)
Max − Min
Min = The smallest value in every column
Max = The largest value in every column.
5.2.1 Scaling to keep the physical properties

The scaling method that keeps the physical properties of the variables, which
was introduced in Chapter 4, is also used to classify with SVM. It is here
applied to the linear scaling. First, the voltage V and I are calculated in p.u.
[13] as
V I
Vp.u. = and Ip.u. = ,
Vbase Ibase
with a common system base Sbase . Which results in,
Sbase
Ibase = √ .
3Vbase
Then the variables are divided into groups e.g. busbar V, load P, line I and
angle differences. The groups are scaled with the linear function (5.11).
Instead of finding Min and Max in every column, the Min and Max values
are found in groups of variables which are then scaled group-wise.
5.2.2 Scaling of unit stator current after capacity

The capacity of the power system to fulfil the (n − 1) criterion depends on
the spinning reserve. Therefore this is used to improve the classification.
Instead of using the unit stator currents in the data set the stator current is
used to calculate the unit capacity to increase the production. When the
stator current is used in the data set no consideration of the unit capacity is
taken. If a unit is idle, the stator current is almost 0. The unit has a large
capacity to prevent a stressed situation. When a unit is not connected to the
power system the stator current is 0, but the unit has no capacity to prevent
a stressed situation. Both units have almost the same stator current, but the
capacity to prevent a stressed situation is totaly different. Instead of using
Ip.u. , I∗ is calculated as
Irated − I
I∗ = ,
Ibase
were Irated is on machine base. Using I∗ the influence of the hyperplane
from a unit in nominal production and a disconnected unit are the same.
5.2.3 Variables
In total 653 variables are used in an example to describe the operating point
of the power system, post-fault. The variables are
• P, Q and I for lines, (156)
• V, f for busbars, (64)
• P, Q and I for loads, (180)
• P, Q and I for transformers, (24)
• P, Q and I for units, (108)
• Q and I for shunts, (84)

5.3. SVM training 45
• Angle differences, (37),
where the number of variables in each group are shown in parenthesis.
5.3 SVM training
Using the OSU SVM matlab toolbox [27] the support vectors are calculated.
The number of support vectors is 20. The support vectors are used with (5.9)
to calculate the hyperplane. The hyperplane is the optimal linear separation
between the two classes. Using (5.4), the function value for each of the
examples in the training set is decided.
In Figure 5.3 (a)-(c) the function value for the 30 examples in the training
set are marked by a ”o” or ”*” (”o” has the binary code -1 and ”*” +1). The
examples which are used as support vectors have the function value ±1, as
in Figure 5.3 (a)-(c). It is possible to find a hyperplane which separates the
two classes in the training set, scaled by 3 different methods, with SVM.
In Section 3.3 one of the traditional methods to indicate if the (n − 1) is
fulfilled in the Swedish part of the Nordel system, the cut-sets methodology
was introduced. Figure 5.4 shows that it is difficult to use the cut-sets to
classify the stability of the examples in the training set. The examples with
different binary code can not be separated with a certain threshold.
As the number of available examples is small, the performance of the
training set is validated by leave-one-out cross validation, see [38]. Each
time 29 of the 30 examples in the training set are used to train the SVM.
The remaining example is then used for classification. This is done for all 30
examples, see Figure 5.3 (d). With the leave-one-out test 28 of 30 examples
in the training set are correctly classified. The robustness in the training set
is pretty good with a fault classification of 28/30.
3 3
2 2
1 1
Function value
Function value
0 0
−1 −1
−2 −2
−3 −3
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
(a) (b)
3 3
2 2
1 1
Function value
Function value
0 0
−1 −1
−2 −2
−3 −3
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
(c) (d)
Figure 5.3: Linear SVM with soft-margin optimization, where C = 1000. Ex-
amples marked with ”o” has the binary code -1 and ”*” has the
binary code +1. The function value, see (5.4), for the training set.
All 633 feature variables are used, 20 variables are excluded due
to variance equal to 0. (a) The variables are scaled linear group-
wise, the unit stator current are also scaled after capacity. (b) The
variables scaled linear group-wise. (c) The variables scaled lin-
ear separately. (d) Leave-one-out test. The function value, for the
training set, 28 of 30 examples are correctly classified. The vari-
ables are scaled as in (b).
5.4. Classifying with a hyperplane 47
3500 2000
1800
3000
1600
2500 1400
1200
Cut−set value
Cut−set value
2000
1000
1500
800
1000 600
400
500
200
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
(a) (b)
Figure 5.4: (a) The sum of transferred active power through Snitt2, for the
examples in the training set. (b) Snitt4, for the examples in the
training set.
5.4 Classifying with a hyperplane
The hyperplane described in Section 5.3 is used to classify the examples in

the classification set. With Equation (5.4) the function value is decided. The
classification set is scaled with the three methods, which corresponds to the
examples in the training set. In Figure 5.5 (a)-(c) the function value, for the
examples in the classification set are marked as in Figure 5.3.
The four examples that are misclassified are classified to fulfill the (n − 1)
criterion, but they do not. The examples are originally represented in the
training set and there they fulfill (n − 1), but in the classification set the
active power load is increased on a common system base so they do not. The
faults which make the examples to not fulfill the (n − 1) are short circuit of
busbar CT11_ B400, CT21_ A400 and FT50_ A400, see Figure 3.2. These
faults also cause blackouts for the examples in the training set so the faults
are represented both in the training and classification set.
The misclassification depends on that the training set has too few ex-
amples, or that the variability of the examples in the training set does not
describe the examples in the classification set. If the training set does not
describe the examples in the classification set it is impossible to make a
correct classification. The number of variables used to describe the clas-
sification problem can also be too large, thus the hyperplane is over-fitted,
see Figure 5.6. The hyperplane is over-fitted if the hyperplane is adjusted to
the training set instead of the classification problem, which hopefully is de-
scribed by the training set. The linear SVM decreases the risk for overfitting,
compared to using a nonlinear function. There is a need to investigate if the
classification can be improved with fewer variables.
As with the examples in the training set the cut-set methodology is not
successful to use to classify the examples in the classification set, see Fig-
ure 5.7. The examples with different binary code can not be separated with
a threshold.
5.4. Classifying with a hyperplane 49
3 3
2 2
1 1
Function value
Function value
0 0
−1 −1
−2 −2
−3 −3
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example
(a) (b)
20
15
10
5
Function value
−5
−10
−15
−20
0 2 4 6 8 10 12 14 16
Example
(c)
Figure 5.5: Linear SVM classification with soft-margin optimization, where

C = 1000. Examples marked with ”o” has the binary code -1 and
”*” has the binary code +1. (a) The function value, for the clas-
sification set. All 633 feature variables are used and scaled linear
group-wise, the unit stator current are also scaled after capacity .
(b) The function value, for the classification set. All 633 feature
variables, scaled linear group-wise. (c) The function value, for the
training set. All 633 features variables scaled linear separately.
Figure 5.6: Illustration of overfitting classification.
3500 2000
1800
3000
1600
2500 1400
1200
Cut−set value
Cut−set value
2000
1000
1500
800
1000 600
400
500
200
0 0
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example
(a) (b)
Figure 5.7: (a) Snitt2, for the classification set. (b) Snitt4, for the classification
set.
5.5. Reducing variable sets 51
5.5 Reducing variable sets
There is a need to analyze if the original data set consists of variables which
cause over-fitting problems. The problem is, given the original data set in
Rd , what is the best subset of feature variables in R c , where c<d? It is not
possible to find an optimal solution for this problem. First of all, we do not
know the dimension of Rc and if we knew it, the number of possible subsets
is
d!
nc =
(d − c)!c!
which is very large when d is 650.
Therefore, a new method to reduce the dimension of the data set is intro-
duced, thus reducing the dimension of the data set without removing vari-
ables which are important for the separation. First, all variables in the train-
ing data with zero variance are removed. Then a method for further reduction
is introduced. To reduce the variables the whole training data is scaled ac-
cording to (5.11), and a hyperplane is calculated with SVM. By guidance
from the normal w of the hyperplane, the dimension of the training set is
reduced.
The unit normal of the hyperplane wn is calculated, as
w
wn = . (5.12)
||w||
The absolute value of the j’th component in w n indicates the contribution

of the j’th variable to the hyperplane, see Figure 5.8. All components lower
than a certain threshold, heuristically found, are removed. This procedure is
important to analyze, so that over-fitting problems can be avoided.
The reduction method is applied on the data set, where the variables are
reduced by guidance from the normal to the hyperplane. After the reduction a
new hyperplane is calculated with the new variable setup. Then the function
value for the examples in the classification set is calculated.
Only 2 of 16 are misclassified, both with the original variable setup and
when the number of variables is reduced from 633 to 70, see Figure 5.9. The
variables that are used to describe the classification problem after reduction
are,
e
j
Figure 5.8: Reducing variables by guidance form the normal of the hyper-
plane.
• P, Q and I for lines, (0, 9, 16 of 156)
• V, f for busbars, (5, 0 of 64)
• P, Q and I for loads, (0, 0, 0 of 180)
• P, Q and I for transformers, (0, 0, 0 of 24)
• P, Q and I for units, (3, 6, 1 of 108)
• Q and I for shunts, (11, 16 of 84)
• Angle differences, (2 of 37),
where the numbers of variables in each group is shown, ordered from left to
right, in parenthesis.
5.5.1 Different variable setup

In [11], [14] and [28] it is indicated that variable have different variations pat-
terns. Consideration must be taken about the difference between variables.
3 3
2 2
1 1
Function value
Function value
0 0
−1 −1
−2 −2
−3 −3
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example
(a) (b)
Figure 5.9: Linear SVM classification with soft-margin optimization, where

C = 1000. Examples marked with ”o” has the binary code -1 and
”*” has the binary code +1. The variables are scaled linear group-
wise, the unit stator current are also scaled after capacity. (a) All
633 feature variables. (b) The 70 variables which have the largest
influence on the hyperplane.
The variables can be divided into two groups:

• Qualitative
• Quantitative
Qualitative variables are normally represented numerically by codes. The
easiest representation is a single binary code as 0 or 1, e.g. circuit breakers
status. In this study circuit breakers/disconectors are not used in the data set,
but shunt capacitors and reactors are included in the data set and they are
approximately qualitative.
Quantitative variables may assume any numerical value along a continu-
ous scale. All variables in the data set are of this type except shunts. It is
important to be aware of the difference between the two variable types.
There is a need to investigate if the number of variables that are used for
classification can be reduced. Large data sets can cause over-fitting prob-
lems, it is also interesting to minimize the cost to measure and collect the
data set. Therefore the different variable groups are used, one-at-a-time, to
calculate the hyperplane with soft margin optimization. The variables are
scaled group-wise, unit stator current is also scaled after capacity. In Figures
5.10-5.12 the results with the different variables are shown, some variables
give a better hyperplane for separation than others. The load and frequency
gives an unsatisfactory separation while line and voltage gives successful
separation.
Bus voltage P load

30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
Q load I load
30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
Q shunt I shunt
30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
Figure 5.10: The function value for the SVM training based on the different
variable groups in the training set. The soft-margin optimization
is used to calculate a hyperplane with slack variables, where C =
1000, for further information see Section 5.1.2. The variables are
scaled group-wise, where the unit stator current also are scaled
after capacity.
P unit Q unit
30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
I unit Angle differences

30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
P line Q line
30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
after capacity.
I line P transformer
30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
Q transformer I transformer
30 30
20 20
10 10
Function value
Function value
0 0
−10 −10
−20 −20
−30 −30
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Example Example
Frequency
30
20
10
Function value
−10
−20
−30
0 5 10 15 20 25 30
Example
after capacity.
1.5
Scaled variable value

0.5
−0.5
−1
−1.5
−2
0 5 10 15 20 25 30
Example
Figure 5.13: Two of the variables in the training set, scaled linear group-wise,
CT22 X1 CurrentValue, qualitative, marked with ”” and CT22
A400 VoltageValue, quantitative, marked with ”+”.
It is possible to calculate a hyperplane that separates the examples in the

training set with both P shunt and I shunt. Therefore, it should be natural
to use the two variable classes for the classification problem. But P shunt
and I shunt are qualitative and therefore they have too large impact on the
hyperplane. The hyperplane will then be trained to classify between two
different shunt patterns that can be identified in the training set. In Table 5.1
the 30 variables which have the largest impact on the hyperplane are listed,
19 of these are shunts. The difference between qualitative and quantitative
variables are shown in Figure 5.13. It is well known that mixing qualitative
and quantitative variables can give problems when building classifiers. There
are methods for dealing with this, e.g. [21] and [38]. Usually these methods
require large training sets.
Table 5.1: The 30 variables which have the largest influence on the hyperplane.
The soft-margin optimization is used to calculate a hyperplane with
slack variables, where C = 1000, for further information see Sec-
tion 5.1.2. The variables are scaled group-wise, where the unit sta-
tor current also are scaled after capacity.
Object w
||w||
CT22 X1 Shunt reactor (I) 0.4675
CT22 X1 Shunt reactor (Q) 0.2635
FT41 EK1 Shunt capacitor (I) 0.2276
FT42 X1 Shunt reactor (I) 0.1483
CT11 T2 Unit (Q) 0.1364
FT41 EK1 Shunt capacitor (Q) 0.1309
CT22 A400 Busbar (V) 0.0871
CL1 Line (I) CT11 0.0867
FT42 X1 Shunt reactor (Q) 0.0837
FT44 T11 Unit (Q) 0.0799
FT51 G2 Unit (Q) 0.0752
CT72 G2 Unit (P) 0.0733
CL8 Line (Q) CT22 0.0727
CT122 A130 Busbar (V) 0.0718
FL14 Line (I) FT51 0.0690
FL15 Line (I) FT51 0.0690
CL3 Line (Q) CT22 0.0671
Due to the different variable types, different variable setups are applied to
classification with SVM. In Section 5.4 the data set was scaled with three
different methods. All three are also used here because no method could be
excluded from the results in Section 3. Five variable setups are compared,
1. Original variable setup, 653
2. Shunts removed, 569
3. All P and Q removed, 289
4. All P and Q, transformer and shunts removed, 257
5. All P and Q, frequency and shunts removed, 225
and the data set is scaled by three different methods,
1. Group-wise
2. Group-wise, unit stator current is also scaled by capacity
3. Separately.
The variable setup which gives the best separation are 2, 4 and 5.
Variable setup 1, scaling method 1 Variable setup 1, scaling method 2

20 20
15 15
10 10
5 5
Function value
Function value
0 0
−5 −5
−10 −10
−15 −15
−20 −20
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example

20 20
15 15
10 10
5 5
Function value
Function value
0 0
−5 −5
−10 −10
−15 −15
−20 −20
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example

20 20
15 15
10 10
5 5
Function value
Function value
0 0
−5 −5
−10 −10
−15 −15
−20 −20
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example
Figure 5.14: The function value for the SVM classification based on the dif-
ferent variable setups and different scaling methods. The soft-
margin optimization is used to calculate a hyperplane with slack
variables, where C = 1000.

20 20
15 15
10 10
5 5
Function value
Function value
0 0
−5 −5
−10 −10
−15 −15
−20 −20
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example

20 20
15 15
10 10
5 5
Function value
Function value
0 0
−5 −5
−10 −10
−15 −15
−20 −20
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example

20 20
15 15
10 10
5 5
Function value
Function value
0 0
−5 −5
−10 −10
−15 −15
−20 −20
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example

20 20
15 15
10 10
5 5
Function value
Function value
0 0
−5 −5
−10 −10
−15 −15
−20 −20
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Example Example
Variable setup 5, scaling method 3

20
15
10
5
Function value
−5
−10
−15
−20
0 2 4 6 8 10 12 14 16
Example
The method to reduce variables by guidance from the normal to the hy-
perplane is applied to the data sets with the variable setups which give the
best classification, according to Figure 5.14-5.16. The variable setups with
shunts are excluded. The linear scaling method is not used. The reduction
is illustrated with a new plot, reduction plot, which shows the number of
correct classified examples in the classification set. The number of variables
in the data set is reduced by two different reduction methods:
1. The first one by increasing the threshold so one variable is excluded at

a time, see Figure 5.17-5.18.
2. The other is by excluding the variable which has the least influence
on the hyperplane, the new data set is used to calculate the difference
between the variables. Thereafter a new hyperplane is calculated and
the process is repeated, see Figure 5.19-5.20.
The variable groups that are dominating in the reduction process are pre-
sented in Table 5.2. The variable setups that have varying classification result
are divided into subsets, see Figure 5.17-5.20. In these subsets the dominat-
ing variable groups are presented in Table 5.2. This is only a guideline of
which variable groups that are excluded, there are other variable groups rep-
resented in the subsets. In Table 5.3 the 30 variables which have the largest
impact on the hyperplane are listed. The variable setup where P, Q, trans-
former and shunt are reduced is used. These 30 variables are also pointed
out in Figure 5.21.
Variable setup 2, scaling method 1, reduction method 1 Variable setup 2, scaling method 2, reduction method 1
16 16
14 14
12 3 2 1 12
Correct classifications
3 2 1
10 10
4
8 8 4
6 6
4 4
2 2
0 0
0 100 200 300 400 500 0 100 200 300 400 500
Number of variables Number of variables
(a) (b)
16 16
14 14
3 2 1 2 1
12 12
10 10
8 8
4 3
6 6
4 4
2 2
0 0
0 50 100 150 200 250 0 50 100 150 200 250
(c) (d)
Figure 5.17: The reduction plot based on the different variable setups and dif-
ferent scaling methods. The x-axis show the numbers of variables
that are used in the classification. The y-axis show the number
of correct classified examples in the classification set. The soft-
16 16
14 14
12 12
3 2 1 2 1
10 10
8 8
6 6 3
4
4 4
2 2
0 0
0 50 100 150 200 0 50 100 150 200
(a) (b)
16 16
14 14
12 12
10 10
8 8
6 6
4 4
2 2
0 0
0 100 200 300 400 500 0 100 200 300 400 500
(a) (b)
16 16
14 14
12 12
10 10
8 8
6 6
4 4
2 2
0 0
0 50 100 150 200 250 0 50 100 150 200 250
(c) (d)
16 16
14 14
12 12
4 3 2 1 3 2 1
10 10
8 8
6 6
4 4
2 2
0 0
0 50 100 150 200 0 50 100 150 200
(a) (b)
Table 5.2: The dominating variable groups that are excluded in the reduction
in Figure 5.17-5.20
1 2 3 4
5.17 (a) Load (P, Q) Load (I) Mixture of Line (P, Q, I)
Angle differ- all variables Unit (P, Q I)
ence Busbar (V)
5.17 (b) Load (P, Q) Load (I) Mixture of Line (P, Q, I)
Line (P) all variables Unit (P, Q I)
Busbar (V)
5.17 (c) Load (I) Load (I) Line (I) Line (I)
Busbar (f) Unit (I) Angle differ- Busbar (V)
ence Angle differ-
ence
5.17 (d) Load (I) Line (I) Busbar (V)
Busbar (f) Unit (I) Line (I)
Angle differ-
ence
5.18 (a) Load (I) Line (I) Line (I) Line (I)
Load (I) Angle differ- Busbar (V)
Angle difference Unit (I)
ence Busbar (V)
5.18 (b) Angle differ- Line (I) Line (I)
ence Busbar (V) Busbar (V)
Line (I) Unit (I)
Load (I)
5.20 (a) Load (I) Load (I) Angle differ- Line (I)
ence Busbar (V)
Line (I)
Unit (I)
5.20 (b) Mixture of Mixture of Busbar (V)
all variables all variables
Table 5.3: The 30 variables which have the largest influence on the hyperplane
in Figure 5.17, variable setup 4, scaling method 2 and reduction
method 1.
Object w
||w||
CT11 A400 Busbar (V) 0.3245
FT44 A400 Busbar (V) 0.3212
CT22 A400 Busbar (V) 0.3100
CT111 A130 Busbar (V) 0.2843
CT122 A130 Busbar (V) 0.2272
FL2 Line (I) FT42 0.2249
FL6 Line (I) FT61 0.1983
FT41 A400 Busbar (V) 0.1708
CT21 A400 Busbar (V) 0.1678
CL1 Line (I) CT11 0.1656
CL5 Line (I) CT12 0.1560
CT12 A400 Busbar (V) 0.1474
CT72 G2 Unit (I) 0.1423
FT41 G1 Unit (I) 0.1260
CT72 G3 Unit (I) 0.1201
AT111 A130 Busbar (V) 0.1189
CL4 Line (I) CT12 0.1121
CT112 A130 Busbar (V) 0.1071
CL7 Line (I) CT71 0.1065
CL6 Line (I) CT71 0.1065
FL3 Line (I) FT43 0.1047
FT42 A400 Busbar (V) 0.0956
FT142 A130 Busbar (V) 0.0937
CT11 T1 Unit (I) 0.0913
FL10 Line (I) FT62 0.0889
CL17 Line (I) FT44 0.0878
FL15 Line (I) FT51 0.0855
FL14 Line (I) FT51 0.0855
FT50 A400 Angle FT62 A400 0.0852
RT132 G1 Unit (I) 0.0851
Figure 5.21: The NORDIC32 system with the variables from Table 5.3
marked. Bus voltage are marked with ”V”, unit stator current
with ”U” and line current with ”—”.
5.6 Implementation
It is not a problem to calculate a hyperplane which can be used to classify the

stability of the NORDEL power system. The calculation time for a training
set which consists of 12000 examples and 1200 variables takes 3,3 s 1 and to
decide the function value for an example with Equation 5.4 takes 0,04 s. The
problem is to obtain a training set which is representative for the NORDEL
system. The SCADA system gives the possibility to measure and collect the
data. But for each of the examples, which can be measured and collected,
it is impossible to decide the binary code. An example can obviously be
modelled in a simulator, for example ARISTO, and the binary code can be
decided for the model. But it is important to be aware of the difference
between model and reality. The questions that have to be investigated before
implementing this technique are:
• Is it possible to produce a training set with a model of the power sys-
tem and use the training set to classify the stability of the real power
system?
• Is it possible to classify an example, without having verified that the
training set is representative for the classification example?
• What happens when a part of the data is misleading and contains mea-
surement errors?
The author considers that it is possible to produce a training set in a sim-
ulator, it is mainly based on the methods which are used today in the state
estimators. But the training set must be representative for the power system
and the amount of time to produce such a training set will be very large. In
[39] the system that have been used consists of 561 buses, 1000 lines and 61
generators. The number of examples in the data base are 11 000. The sec-
ond question inspires for further research in this area, especially in finding
methods to verify if it is possible to classify an example with the existing
training set. The last question is not a large problem if the misleading data
is identified. Robustness studies must be made of which variables that can
be excluded. If the misleading data can be excluded the calculation time to
remove these variables and calculate a new hyperplane is not a problem.
1
On a 1,6 GHz Intel centrino PC with 512 MB RAM-memory and Windows XP using
Matlab 6.5.
5.7. Results 73
5.7 Results
Classification of the stability of the NORDIC32 system with a linear SVM

shows that it is not enough to use the cut-sets methodology as indication
of instability, which is mainly used in the Nordic power system today, see
Figure 5.4 and 5.7. The examples with different binary code can not be
separated using the power flow through the cut-sets.
In Section 5.3 it is shown that a hyperplane can be calculated that separates
the examples with different binary code. But due to the misclassifications in
the classification set, see Figure 5.5, different variable setups, scaling meth-
ods and a reduction method to improve the classification are used. In Figure
5.12 the different groups of variables are used one-at-a-time to calculate a
hyperplane with soft-margin SVM. Some variable groups are more suitable
to use for the classification than others. The difference between qualitative
and quantitative variables must be noticed. In Table 5.1 it can be seen that
18 of the 30 variables that have the largest influence on the hyperplane are
qualitative. It is reasonable to assume that the qualitative variables contribute
to an over-fitted hyperplane which classifies the disconnection of shunt re-
actors and capacitors instead of the stability of the power system, see Figure
5.13. When the qualitative variables are excluded from the data set the clas-
sification is improved. The number of correct classification and the distance
between the two classes is increased.
In this chapter a new method to scale variables which keeps the physi-
cal properties of the system by group-wise scaling is introduced, the unit
stator currents are also scaled by capacity. Furthermore, a new method to
reduce the number of feature variables is also introduced. In Figure 5.17-
5.20 the different variable setups, scaling methods and reduction methods
are validated by the reduction plot. It can be seen in Figure 5.17-5.20 that
the group-wise scaling method with unit stator current scaled after capacity
gives the best classification. The variables that should be reduced from the
data set are, P, Q, transformer and shunts. In Table 5.2 -5.3 and Figure 5.21
it is shown that the variables that have most influence on the hyperplane are:
• Busbar voltage magnitude.
• Line current magnitude.
• Generator current magnitude.

5.8 Conclusions
In this chapter support vector machines are used to calculate a hyperplane

to classify the stability of power systems. It is shown that it is possible to
calculate a hyperplane which separates the two classes. The unique feature
from SVM, which calculate a hyperplane with the maximum distance be-
tween the two classes, gives new possibilities to improve the classification of
power system stability. The classification can be improved by using a vari-
able setup where all P and Q, shunts and transformers are excluded. With the
scaling method which keeps the physical properties of the variables and the
capacity of the units, the classification is improved. The number of variables
can furthermore be reduced by guidance from the normal of the hyperplane.
The data set must be selected with knowledge of the process, using al-
ready known information and also a great consideration of what we want to
classify. In [11] it is mentioned that all variables that can be used should be
used, because they have a contribution to the description of the process. In
practice it is often a large covariance between variables and it can cause a
fault classification due to over-fitting classifications models [14].
The dataset that is used to improve the classification of power system sta-
bility with SVM has not the proper variability for a final implementation.
With a proper variability of the training set the corresponding support vec-
tors can be used for development of an early warning system, in real-time.
This application can be developed in Simulink interacting with the ARISTO
simulator.
Chapter 6
Conclusions
All results indicate that SVM is an effective method for classification of

power system stability. The thesis presents some of the issues that need to
be adjusted to improve the classification result. The work involves the use of
statistical learning algorithms to solve problems in power system operation.
This thesis should also inspire for future research, because there is a large
amount of problems that have not been solved yet.
6.1 Summary of the results
The SVM method is more adaptable to the classification problem in this the-
sis than PCA and the result is much more satisfying. With the linear SVM
the hyperplane is calculated that separates the stable from the unstable ex-
ample in the training set with the maximum distance between them. The
hyperplane gives a better indication of the security of a state than the cut-set
methodology. The calculation time is also short due to the global optimiza-
tion problem of the SVM. In general, the SVM does not perform well outside
the span of the training data. The training and classification set in this thesis
are too small. In the work by Wehenkel, see [39], the number of examples in
the training set is 11 000. For further work the number of examples must be
increased.
The three main contribution of this thesis are:
• The method to scale the variables to keep the physical properties and
the capability of the units, see Section 5.2.
75
76 Chapter 6. Conclusions
• The method which excludes the binary variables, due to the difficulties
to classification with SVM with a mixture of binary and continuous
variables, see Section 5.5.1.
• The method to reduce the number of variables by guidance from the

hyperplane, see Section 5.5.
These three methods are introduced in this thesis and they contribute to im-
prove the classification result.
6.2 Future work
The classification result in this thesis is quite successful, but the method must
be further verified before a possible implementation in power system opera-
tion. The size of the training and classification sets are too small and must
be increased. With the ARISTO simulator it is possible to produce realistic
synthetic data, thanks to the detailed models and the dynamic simulations.
However, the interface of ARISTO is not developed for this type of study.
The amount of time it takes to develop the data set is very long. Therefore
consideration must be taken if ARISTO should be used in the future for this
project. The dimension of the training set should be at least the same number
of examples as measured variables, 653.
In this thesis linear SVM was used and the impact of adding non linear
functions to the hyperplane has not been investigated. Therefore an analysis
of the effect of adding non linearities to the SVM to solve the classification
problem is motivated. The non linear SVM provide the user to the choose
what non linear functions that shall be added and also the order of them. The
frequently used functions today are polynomial, sigmoid and gaussian.
The scaling function was modified in this thesis to compensate for the
capability of the generators. It should be interesting to develop methods
which also compensate for the transmission capability of the power system.
By adding this, the changes in the transmission, e.g. disconnection of lines
and changes in transmission capability due to over heated lines, could easier
be observed by the separating hyperplane. The improvements by adjusting
the scaling method will hopefully make the training set more robust.
In this thesis no methods have been developed for dealing with the prob-
lem of how to check if the training set is representative for the classification
6.2. Future work 77
example. In general, classification methods can not be expected to perform

well outside the span of the training data. With a larger training set, it would
be interesting to study the structure and span of the training data in the space
of the measured variables.
78 Chapter 6. Conclusions
Bibliography
[1] Nordel grid planning rules for the nordic transmission system.
www.nordel.org, 2005-02-27, 1992.
[2] Rekommandasjon for frekvens, tidsavvik, reglerstyrke og reserve.

www.nordel.org, 2005-02-27, 1996.
[3] Elavbrottet 23 september 2003 - händelser och åtgärder.

www.nordel.org, 2005-04-25, 2003.
[4] http://www.nordel.org, 2005.
[5] CIGRE TF 38.02.08. Long term dynamics. phase ii. final report. Tech-
nical report, CE/SC 38 GT/WG 02, Ref. No. 102, 1995.
[6] S. Arnborg. Emergency control of power systems in voltage unstable

condition. PhD thesis, Royal Institute of Technology, Sweden, 1997.
[7] H. Blomqvist, editor. Elkraftsystem 1. Liber AB, 1997. ISBN 91-47-

00064-3.
[8] N. Cristianini and J. Shawe-Taylor. An introduction to support vector

machines and other kernel-based learning methods. Cambridge Uni-
versity Press, 2000. ISBN 0-521-78019-5.
[9] A. Edström and K. Walve. The training simulator aristo - design and ex-
periences. In IEEE Power Engineering Society Winter Meeting, 1999.
[10] L. Eriksson, E. Johansson, N. Kettaneh-Wold, C. Wikström, and

S. Wold. Design of experiments principles and applications. Umet-
rics, 2000.
79
80 Bibliography
[11] L. Eriksson, E. Johansson, N. Kettaneh-Wold, and S. Wold. Multi- and

megavariate data analysis. Umetrics, 2001. ISBN 91-973730-1-X.
[12] L. H. Flink and K. Carlsen. Operating under stress and strain. In IEEE
Spectrum 15, 1978.
[13] J. Glover and M. Sarma. Power system analysis and design.

Brooks/Cole, 2002. ISBN 0-534-95367-0.
[14] T. Hastie, R. Tibshirani, and J. Friedman. The element of statistical

learning; data mining, inference and prediction. Springer Series in
Statistics, 2001. ISBN 0-387-95284-5.
[15] A. Höskuldsson. Prediction methods in science and technology. Thor

Publishing, 1996. ISBN 87-985941-0-9.
[16] K. Hotta. Support vector machines with local summation kernel for
robust face recognition. In Proceedings 17th International Conferance
on Pattern Recognition, 2004.
[17] C-W. Hsu, C-C. Chang, and C-J Lin. A practical guide to support vector
classification. Technical report, Department of Computer Science and
Information Engineering, National Taiwan University, 2002.
[18] T. Joachims. Text categorization with support vector machines: learn-

ing with many relevant features. In Proceedings of ECML-98, 10th
European Conference on Machine Learning, 1998.
[19] B. Johansson. Computer vision using rich features - geometry and sys-
tems. PhD thesis, Lund Institute of Technology, 2002.
[20] R. Johnson and D. Wichern. Applied multivariate statistical analysis.

Prentice Hall, 2002. ISBN 0-13-834194-X.
[21] W. J. Krzanowski. Discrimination and classification using both binary

and continuous variables. Journal of the American statistical associa-
tion, 1975.
[22] P. Kundur. Power system stability and control. McGraw-Hill, Inc, 1993.
ISBN 0-07-035958-X.
Bibliography 81
[23] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose,

C. Canizares, N. Hatziargyriou, D. Hill, A. Stankovic, C. Taylor, T. Van
Cutsem, and V. Vittal. Definition and classification of power system
stability. IEEE Transaction on Power Systems, 19(2), 2004.
[24] T. Larsson. Driftsinstruktion, spänningsreglering på stamnätet. Tech-

nical report, Svenska Kraftnät, 2003.
[25] G. Lindgren and H. Rootzén. Stationära stokastiska processer. KFS i

Lund, 2003.
[26] P-A. Löf. On static analysis of long-term voltage stability. PhD thesis,
Royal Institute of Technology, Sweden, 1995.
[27] J. Ma, Y. Zhao, and S. Ahalt. Osu svm classifier matlab toolbox (ver
3.00), 2002. web http://eewww.eng.ohio-state.edu/ ∼maj/osu_svm.
[28] J. N. Miller and J. C. Miller. Statistics and chemometrics for analytical

chemistry. Prentice Hall, 2000. ISBN 0-130-022888-5.
[29] D. Montgomery. Design and analysis of experiments. John Wiley &

Sons, 2001. 0-471-73304-0.
[30] L. S. Moulin, A. P. Alvares da Silva, M. A. El-Sharkawi, and

R. J. Marks II. Support vector machines for transient stability analysis
of large-scale power systems. IEEE Transaction on Power Systems, (2),
2004.
[31] E. Osuna, R. Freund, and F. Girosi. Training support vector machines:

an application to face detection. In Proceedings Conference Computer
Vision and Pattern Recognition, 1997.
[32] K. Pearson. On lines and planes of closest fit to systems of points in

space. In J. Edcu. Psych., volume 2, pages 559–572. 1901.
[33] B. D. Ripley. Pattern recognition and neural networks. Cambridge

University Press, 1996. ISBN 0-521-46086-7.
[34] C. Taylor. Power system voltage stability. McGraw-Hill, Inc, 1994.

ISBN 0-07-063184-0.
82 Bibliography
[35] E. Vaahedi, C. Fuchs, Y. Mansour W. Xu, H. Hamadanizadeh, and

K. Morison. Voltage stability contingency screening and ranking. IEEE
Transaction on Power Systems, 14(1), 1999.
[36] V. Vapnik. The nature of statistical learning theory. Springer Verlag,

1995. ISBN 0-387-98780-0.
[37] V. Vapnik, B. Boser, and I. Guyon. A training algorithm for optimal

margin classifier. In Proceedings of the 5th Annual ACM Workshop on
Computer Learning Theory, 1992.
[38] A. Webb. Statistical Pattern Recognition. John Wiley & Sons, 2002.
ISBN 0-470-84513-9.
[39] L. Wehenkel. Automatic learning techniques in power systems. Kluwer

Academic Publisher, 1998. ISBN 0-7923-8068-1.

Power System Security Assessment: Application of Learning Algorithms

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Power System Security Assessment: Application of Learning Algorithms

Uploaded by

Copyright:

Available Formats

Power System Security Assessment

Application of Learning Algorithms

c Christian Andersson, 2005

2 Power system control 9

4 Principal component analysis 27

5 Support vector machines 37

First of all, I would like to thank my supervisors, Bo Eliasson for introducing

Malmö, August 21, 2005

In this thesis classification of power system stability defines a method to

1.1 Previous work

1.2 Objectives and contributions of the thesis

The main objective of this thesis is to investigate if it is possible to use sta-

• Find a method which makes it possible to classify the stability of a

• Adjust the classification method for the application to power system

• Investigate if the classification method can be implemented in real-

• The power system stability is classified with a hyperplane, which is

• Methods to improve the classification by using an appropriate scaling

• Methods to reduce the number of measured variables are introduced.

• Method for handling the difficulties of classification with statistical

1.3 Outline of the thesis

The thesis is based on the following publications:

[2] L. Lindgren, B. Eliasson and C. Andersson. Damping of power system

[3] B. Eliasson and C. Andersson. New selective control strategy of power

[4] B. Eliasson and C. Andersson. New selective control strategy of the

Power system control

2.1 Power system stability

Rotor Angle Frequency Voltage

Small-Disturbance Small-Disturbance Large-Disturbance Small-Disturbance

Figure 2.1: Classification of power system stability. Adapted from [23] .

Partial or total Overloads

Control and / or protective actions

Foreseen or unforeseen disturbance

2.2 Power system operation

In power system control today the operating point is validated by a state

system. It is also common to use load-flow calculations. In a load-flow

3.1 Data set structure

3.2 The NORDEL power system

In this presentation of NORDEL, Iceland and Jutland (Denmark) are ex-

Table 3.2: Installed capacity in NORDEL on 31 Dec. 2003, MW.

• Line distance protection

• Line over-current protection

Table 3.3: Installed capacity in NORDIC32, MW.

• Unit over/under frequency protection

• Unit over/under voltage protection

and the automatic equipment

• Load disconnection due to under-frequency

• Shunt capacitor/reactor connect/disconnect due to under/over voltage.

The detailed representation of power system models in ARISTO results in

Figure 3.2: The NORDIC32 system.

3.4 The 2k factorial design

Factorial designs are often used in designs of experiments. The purpose

• Load disconnection from system protection.

Figure 3.3: Illustration of a factorial 23 design.

with a limited number of examples. Therefore factorial design is not used

3.5 Dimension criterion

• There is no load shedding, due to under frequency.

• The final post-fault voltage profile is acceptable.

• The final post-fault frequency is acceptable.

The contingencies are e.g. disconnection or short-circuit of a line, load,

To verify the (n−1) criterion for the examples, a three-phase short-circuit

1. Examples where a single fault leads to system black-out.

2. Examples where a single fault leads to black-out in two or three sub-

3. Examples which fulfill the (n − 1) criterion.